Pricing with causal ML: elasticity, cannibalization and strategy
Causal ML pricing estimates what a price change will cause, using own-price and cross-price elasticity, then sets prices against a stated strategic goal.
By the Snowfox AI team
Price optimization with causal ML means estimating what a price change will cause, not what has historically moved together with price, and then choosing prices across the portfolio that serve a stated business goal. It rests on three estimates: how demand for a product responds to its own price, how it responds to the prices of related products, and where each price sits against competitors. Without a causal estimate of those effects, a pricing model mostly repeats the patterns of past pricing decisions.
Causal ML (causal machine learning) is the family of methods that use machine learning to estimate cause and effect from data, instead of predicting from correlation alone.
Why does correlation-based pricing mislead?
Historical prices were not set at random. Someone lowered a price because demand was weak, raised it before a holiday, ran a promotion on a product that was already popular, or matched a competitor. A model trained on that history learns the pricing team’s habits mixed with the customer’s response, and cannot tell them apart.
Typical ways this goes wrong:
- Promotions coincide with flyers, better shelf positions and seasonal peaks. The model credits the whole jump in sales to the discount and recommends more discounts.
- Prices go up when demand is already high. The data shows high prices next to high volume, and a naive model concludes that customers barely react to price.
The technical name is confounding: a third factor drives both the price and the sales. A forecasting model can be accurate on past data and still be wrong about what happens when you change a price.
The cleanest way to isolate the effect of price is a controlled price test. Where testing everything is not possible, the model adjusts for the known drivers of both price and demand (season, promotion calendar, stock, store profile) and learns from the variation that remains.
What is the difference between own-price and cross-price elasticity?
Elasticity measures how sensitive demand is to price.
| Term | What it measures | Why it matters |
|---|---|---|
| Own-price elasticity | Change in a product’s sales volume when its own price changes | Shows which products can carry a higher price and which lose volume quickly |
| Cross-price elasticity | Change in a product’s sales volume when another product’s price changes | Shows substitutes and complements, so you can price the portfolio and not single items |
An own-price elasticity of -2 means that a 1% price increase reduces units sold by about 2%. A product like that is elastic, and a higher price costs volume fast. A product with elasticity close to zero is inelastic and has room for margin.
In cross-price elasticity the sign carries the meaning. Positive means substitutes: raise the price of one coffee brand and the other brand sells more. Negative means complements: discount pasta and pasta sauce sells more too.
Elasticity is not fixed. It varies by store, customer profile, season and price range, so estimate it at the level where prices are actually set.
What is cannibalization and why does it change the answer?
Cannibalization happens when a gain on one product comes out of another product in your own portfolio. A discount on the house-brand olive oil lifts its sales, but part of that volume used to go to a premium brand with a higher margin. Product by product, the promotion looks like a success. At category level, margin fell.
This is why one product at a time is the wrong unit of work. The optimization has to run over the category or the portfolio, using cross-price elasticities, so that the recommended set of prices is good in total.
Where does competitor position fit?
Competitor position is the price gap between you and the relevant competitors on each product. It matters in two ways. First, it changes demand: shoppers tend to be more price-sensitive on the items they compare. Second, it shapes price image, meaning how expensive customers believe the store is overall.
A model that ignores competitors finds easy margin on exactly the products where an increase is most visible.
The opposite error: do not automate competitor matching and call it optimization. Following every competitor move hands your pricing decision to someone else’s strategy.
How do you align prices to a strategic goal?
An optimizer has to be told what “best” means. That is a business decision, and the model cannot make it for you.
| Goal | What the optimizer does | Typical constraint |
|---|---|---|
| Margin | Raises prices where demand is inelastic and cannibalization is low | A limit on the gap against competitors on visible items |
| Volume or share | Lowers prices where demand is elastic and the extra volume is new, not cannibalized | A margin floor per category |
| Positioning | Holds defined gaps against competitors or between product tiers | Margin and volume are optimized only inside those gaps |
Most companies want a mix, such as margin growth with volume held above a floor. Write it down before the model runs.
Then add the rules the business already lives with, such as price endings, maximum change per cycle and contractual prices. A recommendation that breaks them will not be applied.
How do you know it worked?
Compare against stores that did not receive the new prices. Sales before and after a change prove very little, because season, weather, competitors and the economy changed too.
Matched comparison stores are stores chosen because they resemble the test stores in what drives sales: size, location profile, assortment and past trend. The test group gets the model’s prices, the comparison group carries on as before, and the difference in the outcome over the same period is the estimate of the effect.
Two rules keep the measurement honest:
- Decide the metric and the period before the test starts.
- Measure at category or store level, so cannibalization is counted.
How did this work for a food retailer?
Snowfox’s published case with a food retailer follows this logic. The retailer needed to find the right price combination across its portfolio while balancing demand and margin. The scope was its premium stores serving higher-income neighborhoods.
The work used priceGURU, Snowfox’s price optimization product. A causal model combined demand forecasts, cross-elasticity estimates and the retailer’s market strategy to adjust prices.
The published case reports a 6% margin increase compared with matched comparison stores. That is a difference against comparable stores, not a before-and-after number. The case does not publish the size or length of the test, so none is given here.
Where to start
Start with the problem, not the technology. Before any modeling, answer three questions: what is the goal (margin, volume or positioning), which rules must prices respect, and how will the result be measured.
Then check the data. You need a clean history of prices, sales, promotions and stock at product and store level, plus competitor prices where you can get them. If that history is scattered or unreliable, the first project is data engineering, not pricing.
- priceGURU: price optimization across the portfolio.
- Data engineering: for when price and sales history needs work first.
- Contact: bring the pricing problem, and we will tell you whether a model is the right answer.